ALD assisted nanoplasmonic slot waveguide for on-chip enhanced Raman spectroscopy

Recently, the integration of plasmonic structures on photonic platforms has attracted a lot of attention.^1–9 The combination of integrated photonics with plasmonic structures can indeed provide a strong field enhancement over an increased interaction length resulting in a strong light-matter interaction.^10–13 One of the key applications in this context is surface enhanced Raman spectroscopy (SERS). SERS relies on plasmonics enhancement for increasing the electric field strength by many orders of magnitude. Basic SERS substrates are rough gold surfaces that can exhibit large plasmonic enhancement, thanks to tight gaps, but do not offer enough control, homogeneity, and reproducibility of this enhancement. Many SERS substrates have been developed^14–16 providing more reproducible Raman spectra. In parallel to those developments of SERS, Raman spectroscopy enhanced by waveguides was introduced to increase the interaction length between light and analyte while keeping a confined beam capable of collecting the spontaneously generated Raman signal.¹⁷ Strip waveguides and slot waveguides have succeeded to earlier demonstration^18–22 due to the combination of high waveguide enhancement and ease of use. Those photonics waveguides still offer relatively lower signal enhancement than what is possible with SERS; however, they are far superior in terms of integration with light sources,^23,24 filters,²⁵ spectrometers,²⁶ and detectors,²⁷ which could be integrated on a single chip as a full Raman system. It is therefore tempting to bring a plasmonic enhancement to the existing photonics integrated circuits (PICs). This has been investigated both theoretically and experimentally using nanoplasmonic antennas.^28–31 A later study made use of microsphere lithography to pattern nanotriangles on a photonic waveguide that gave a better SERS enhancement.³² However, in the previous reports, a Raman background due to the underlying Si₃N₄ dielectric waveguide³³ was limiting the SNR; therefore, other realizations have been reported, such as combining existing PIC with metallic nanoparticles but at the expense of quantitatively reproducible Raman spectra.^34–36 Here we report the fabrication of metallic slot waveguides that offer both a strong field enhancement due to the highly confined propagating surface plasmon polariton (PSPP) mode and a limited spurious background, thanks to the limited field overlap with the dielectric waveguide. Moreover, the fabrication avoids any use of e-beam lithography and rather relies on the atomic layer deposition (ALD) to create the nanometer scale gaps required for a large plasmonic enhancement. This makes our approach compatible with the back-end CMOS fab. Our sensor relies on SiN photonic circuits fabricated by DUV lithography and reactive ion etching in a CMOS fab. A regular strip waveguide is interfaced to a slot waveguide via a dedicated tapering section.³⁷ The dielectric waveguide is then conformally coated via atomic layer deposition of Al₂O₃ to reduce the gap of the slot from 150 nm to 34 nm. Subsequently, gold is deposited via sputtering reducing the gap further down to 14.8 nm. The length of the plasmonic structure is defined by a window over which the gold is deposited directly on the Al₂O₃, while the rest of the gold is removed using metal lift-off. Unlike plasmonic nanoantenna, our plasmonic slot waveguide offers micrometer long interaction length and an off-resonant enhancement, hence a broadband Raman conversion efficiency.^38–40 The local enhancement is, therefore, less localized both spectrally and spatially, resulting in an undistorted Raman response and a reduced risk for any field induced damage to the analyte. In this manuscript, we characterize the most important properties of the sensor: SERS enhancement, interaction length, spurious photon background, and signal stability. The sensor is tested using a monolayer of 4-Nitrothiophenol (NTP) whose possible photoreduction to dimercaptoazobenzene (DMAB) is also investigated. In the end, the performance of our reported sensor is compared with other on-chip SERS substrates.

The geometry of the sensor is depicted in Fig. 1. The plasmonic sensor is preceded by a low loss 400 μm long Si₃N₄–Al₂O₃ slot waveguide acting as an access waveguide. The length of the plasmonic waveguide is set to 100 μm, far larger than the decay length (15 μm). The fundamental TE mode of the Si₃N₄–Al₂O₃ slot waveguide at a pump wavelength λ_p excites the gap plasmonic mode with a coupling efficiency γ_p. This propagating surface plasmon polariton (PSPP) mode is strongly confined in the gap (g) and interacts with the analyte (NTP) uniformly lying over the gold. The pump beam excites the vibrational modes of the NTP molecules to generate the enhanced Stokes signal that mostly remains coupled to the hybrid plasmonic mode (in both contra- and co-propagating directions). This enhanced pump excitation and subsequent Stokes signal generation can be computed numerically to estimate the Raman conversion efficiency η (see supplementary material). The backward collected signal increases until a certain waveguide length and then saturates due to the decay of the pump light. The saturation length is set by the total waveguide loss α_p. The total Stokes signal P_s normalized by the total guided pump power P_o for the device operating in the aforementioned setup can be defined numerically as

where ρ and σ are the molecular density and differential Raman cross section of the analyte. The length factor represents the SERS signal dependence on the plasmonic waveguide length (L_p). To analyze the effect of gold thickness t_gold and ALD alumina thickness $tAl2O3$⁠, a COMSOL Multiphysics mode solver is used to compute the eigenmodes of the hybrid $Si3N4−Al2O3−Au$ waveguide. We consider a Si₃N₄ slot waveguide with a refractive index of $nSi3N4=1.9$⁠, coated with an Al₂O₃ layer whose refractive index is $nAl2O3=1.6$ and a final thin coating of gold whose complex refractive index is given by the Johnson and Christy model. Our simulation assumes only hybrid-TE excitation (maximum |E| along the surface of the sample) as this mode possesses the highest field confinement into the gap. The electrical field strength |E| is plotted in Fig. 1, showing that the field confinement (≈92% for g = 20 nm) is much stronger than in the case of the Si₃N₄ slot waveguide (≈35% for g = 150 nm). This confinement is expected to strongly reduce the spurious background flux of photons originating from the guiding core of the waveguide (Al₂O₃ and Si₃N₄). The test monolayer of NTP is 1-nm thick and selectively binds to gold. Knowing its refractive index^41,n_NTP = 1.62 and the Raman resonance of 1339 cm⁻¹ (ν_s(NO₂)), we can estimate the conversion efficiency η when the pump wavelength is set to 785 nm and the Stokes wavelength is 877 nm. For a 100 μm long plasmonic waveguide, the ratio ζ = P_s/P_o of the generated Stokes power over the incoming pump power is plotted in Fig. 2 for a range of thicknesses $tAl2O3$ and t_gold chosen such that the gap (g) ranges from 150 nm to 10 nm. For simplicity, we keep a constant coupling efficiency, i.e., γ_p = γ_s = 5 dB for all combinations of $tAl2O3$ and t_gold. The plot shows a nearly constant enhancement for a fixed gap size that is a fixed $tAl2O3+tgold$⁠. Of course, this trend does not hold anymore when the gold thickness is so small that the plasmonic effect disappears. For our design, we target a thicker alumina layer (∼58 nm) and rather a thin gold layer (∼13-16 nm) for two reasons. Indeed, the thickness of Al₂O₃ is experimentally better controlled (thanks to ALD), leading to an easier fabrication. Moreover, a lesser amount of gold also reduces the reflection at the transition between the access and plasmonic waveguide. For $tAl2O3=$ 58 nm and t_gold = 13 nm, the plasmonic waveguide losses at the pump (α_p) and Stokes wavelength (α_s) measured using the cut-back method are, respectively, 2.4 and 2.1 dB/μm (see supplementary material). Similarly, the measured γ_p and γ_s are 3.95 and 5.1 dB/facet, respectively. For these values of α_p, α_s, γ_p, and γ_s and a plasmonic waveguide length L_p = 100 μm, ζ calculated using the COMSOL Multiphysics mode solver is 1.96 × 10⁻⁹. The single molecule enhancement factor,⁴² i.e., SMEF = |E(λ_p)|²|E(λ_s)|² is also calculated by assuming a dipole located in the middle of the gap, i.e., SMEF = 1.5 × 10⁷.

The details about the fabrication can be read in Sec. V. As discussed earlier, the pump wavelength λ_p = 785 nm is used for all reported results unless it is mentioned explicitly. The description about the experimental conditions and measurement setup is outlined in Sec. V. Raman spectra are measured from both 100 μm long gold coated and 0.6 mm long bare Si₃N₄ slot waveguides, the latter serving as a reference. The results are presented in Fig. 3. Both chips have been processed so that a monolayer of NTP coats the gold nanostructures. Indeed, no NTP peak can be seen from the reference waveguide (free of gold). The spectrum from the plasmonic waveguide shows not only the NTP Raman modes but also a weaker background than the reference waveguide. This is because the pump beam is strongly attenuated in the plasmonic structure and no further Raman background is generated after it. In the absence of gold, light is much less attenuated and a spurious background is generated after the sensor. It is interesting to study the dependence of the SERS signal on the length of the plasmonic waveguide (L_p) and the length of the access waveguide (L_a). In the first set of experiments, the access waveguide length is kept constant, while the plasmonic waveguide length is varied from 0 to 290 μm. As shown in Fig. 3(b), the signal increases with the waveguide length until 15 μm and then saturates due to the decay of the pump light. Fitting the length factor of Eq. (1) gives α_p = α_s = 3.0 ± 1.2 dB/μm, where the large error bar is due to the limited number of points at short waveguide lengths. This value fits well to the α_p and α_s measured using the cut-back method, i.e., 2.4 dB/μm and 2.1 dB/μm, respectively. This interaction length is rather long for a plasmonic structure, and this obviously contributes to the Raman conversion efficiency. It is worth mentioning that α_p depends on the final plasmonic gap (g), i.e., it increases with decreasing plasmonic gap due to the higher modal interaction with gold.⁴³ In the second set of measurements, the effect of the access waveguide length (L_a) is investigated. As confirmed by Fig. 3(a), the access waveguide does not contribute to the NTP spectrum. However, it does generate a broadband background that can mask the SERS signal of interest if its length is too large.⁴⁴ Therefore 8 different access waveguide lengths L_a from 10 to 410 μm with a step size of 50 μm are measured using the same experimental parameters. As shown in Fig. 3(c), the signal to background ratio decreases with increasing access waveguide length. As a third experiment, we explore the spatial reproducibility of the SERS spectra by measuring the relative standard deviation percentage (RSD%) of the SERS signal of all the Raman modes of NTP for 24 different waveguides from three different chips (eight each). Each spectrum is measured using the same pump power, access waveguide length, CCD settings, and integration time. All Raman spectra measured from different waveguides are reported in supplementary material. As shown in Fig. 3(d), all NTP modes except 1080 and 1339 cm⁻¹ show high stability with RSD% < 5%. This limited variation is mainly attributed to differences in coupling efficiencies, random gold roughness, and optical alignment inaccuracy. The increased variability at the 1080 and 1339 cm⁻¹ modes is attributed to the chemical instability of the NTP exposed to high optical fields and will be discussed hereafter. This SERS signal variation in our plasmonic slot waveguide is comparable to state-of-the-art SERS substrates for free-space excitation at 785 nm pump wavelength.^45,46

While most plasmonic structures used for SERS rely on a localized surface plasmon resonance, our plasmonic slot waveguide uses a gap propagating surface plasmon polariton (PSPP) that offers broadband and off-resonant enhancement. As a consequence, we expect a signal enhancement that has a reduced dependence on the excitation- and emission wavelength. To demonstrate this broadband nature of the field enhancement, we have taken SERS spectra at two different excitation wavelengths. The reported spectra in Fig. 4(a) are recorded at identical pump power and integration time but for a pump wavelength of 632 nm and 785 nm. The spectrum at shorter wavelength is stronger as compared to the one resulting from 785 nm excitation. However, the spurious background level also increases due to the relatively stronger photoluminescence of gold at shorter wavelength.⁴⁷ This result highlights the broadband enhancement of our Raman sensor and the fact that it can be used for any Stokes Raman shift contrasting with the field enhancement in nanoplasmonic antenna where the resonance is limited to ≈100 nm. To the best of our knowledge, this is the first broadband SERS demonstration of an on-chip SERS sensor.

Another common limitation of plasmonic structures for Raman spectroscopy is the maximum optical excitation power $Pomax$ that can be used before degrading the sample/analyte under study. Indeed, this limits the pump power that can be used and therefore the strength of the Raman signal. Due to the relatively low localization of the electric field, our propagating plasmon based sensor is expected to have large $Pomax$⁠. When exposed to too large intensity, NTP molecules start to photoreduce in a form of dimercaptoazobenzene (DMAB). In order to find $Pomax$⁠, a series of spectra has been recorded at different pump powers with 1 s integration time each. The photo-degradation of NTP to dimercaptoazobenzene (DMAB) can be monitored using the SERS intensity at 1339 cm⁻¹ vs P_O. The power level at which I₁₃₃₉ shows nonlinearity against P_O corresponds to the photoreduction of NTP into DMAB. As presented in Fig. 4(b), the 1339 cm⁻¹ peak starts to decay after $Pomax$ = 0.8 mW pump power guided into the access waveguide, while the peaks at 1145 cm⁻¹ due to β(CH) and 1383 and 1443 cm⁻¹ due to ν(N = N) increase in intensity that shows the gradual conversion of NTP to DMAB. The Si₃N₄ background at 400 cm⁻¹ also scales linearly with the pump power confirming no photo-damage to Si₃N₄ at higher powers. A similar observation has been reported before,^48,49 where NTP molecules that bind to a SERS substrate were excited at the pump wavelength of 532 and 632 nm using the free space optics.

Next, we made a comparison between the Raman signal strength obtained through a metal slot waveguide excitation and collection (waveguide excited) versus the excitation and collection through a high NA conventional microscope objective [top-down free space (FS)]. A 2 mW pump power before the objective (Zeiss 100×/0.9 EC Epiplan NEOFLUAR: ∞/0) and 60 s integration time are used for both cases. The results are shown in Fig. 5. First, a reference spectrum is measured through top-down excitation and collection from a position on the sample containing no metal slot waveguide. No NTP peak is seen in the reference spectrum (red curve). Next, a spectrum is recorded from a metal slot waveguide (green curve). The spectrum contains the dominant modes of NTP, i.e., 1110, 1339, and 1573 cm⁻¹. Also, both spectra contain the 520 cm⁻¹ Raman detuning of the underlying silicon substrate. Finally, the light is butt coupled to the same metal slot waveguide preceded by 100 μm reference waveguide. The collected Raman spectrum (blue curve) is shown in Fig. 5. The waveguide excited and collected spectrum shows all Raman modes of NTP. The results show that for a fixed pump power before an objective, the Raman signal strength of a metal slot waveguide is approximately 30 times larger as compared to that using a 0.9 NA objective. This is attributed to the long interaction length of the former case. Also, the SNR (Sig/$BG$⁠) of the former case is approximately 7 times larger than that of the latter case. This comparison is based on the signal counts that correspond to the 1339 cm⁻¹ NTP mode.

Finally, we compare our hybrid plasmonic waveguide based sensor (HPW) with previously published on-chip surface enhanced Raman sensors, i.e., nanobowties (NBs)^31,44 fabricated using e-beam lithography and nanotriangles (NTs)³² fabricated using nanosphere lithography. The comparison is presented in Table I in terms of measured Raman conversion efficiency (ζ = P_s/P_o), number of molecules (N_mol) excited, resonance behavior, and fabrication technique. The experimental value of ζ for our nanoplasmonic sensor that is extracted for the 1339 cm⁻¹ mode of NTP is (1 ± 0.057) × 10⁻⁹. This is 3 orders of magnitude higher than that of nanobowties and an order of magnitude higher than that of nanotriangles. The total number of molecules excited (N_mol) in each case is calculated using A_g × ρ_g where A_g is the gold surface area and ρ_g = 4.4 × 10⁶ molecules/μm² (Ref. 50) is the density of NTP molecules on a gold surface. Another advantage of the reported SERS substrate over others is the ease of fabrication that consists of ALD assisted DUV photolithography, which enables mass scale manufacturing. Next, the broadband nature of SERS enhancement is compared with the other two devices. NB and NT are highly resonant, i.e., FWHM of resonant enhancement ≤100 nm for nanobowties and ≤250 nm for nanotriangles. Our sensor offers a non-resonant enhancement, making the SERS enhancement independent of excited and scattered wavelengths. In the end, we numerically compare the signal strength offered by a hybrid plasmonic waveguide (HPW) based sensor with a standard silicon nitride strip and slot (gap = 150 nm) waveguide^19,39 fabricated using DUV photolithography. Two different cases, i.e., a monolayer of NTP and a bulk Isopropyl Alcohol (IPA) cladding, are considered. The results are presented in Table II (italic). In the case of an NTP monolayer, for the sake of comparison, we assume the same NTP Raman cross section (σ = 1.8 × 10⁻²⁹ cm²/sr) and surface density (ρ = 4.4 × 10⁶ molecules/μm²) for both sensors. However, this is rather an artificial assumption because ρ of a given molecule strongly depends on the surface of the substrate as well as the binding technique. But it serves the purpose of comparing different sensors. For an NTP monolayer, 15 μm long HPW exhibits approximately 14 and 20 times stronger signal as compared to a 4 cm long silicon nitride slot and a 5 cm long strip waveguide. Here, we have considered the coupling loss from an access waveguide to a plasmonic waveguide, i.e., γ_p and γ_s are 3.95 and 5.1 dB/facet, respectively. Similarly, for a bulk analyte, ζ is computed assuming a 5 μm thick IPA as a top cladding (n_IPA = 1.37, λ_p = 785 nm, λ_s = 840 nm, σ = 8.0 × 10⁻³¹ cm²/sr, and ρ = 7.87 × 10²¹ molecules/cm³).³⁹ In spite of the high field enhancement offered by a HPW, ζ for the silicon nitride slot waveguide is still approximately 3 times larger than for a HPW. This is attributed to a long interaction length and large modal overlap with the analyte. Note that in this comparison, the spurious Raman background generated from the waveguide core³³ is not taken into account, which is indeed a major limiting factor in achieving high SNR in the dielectric waveguide based Raman sensors.

In conclusion, we have reported a SERS sensor suitable for on-chip integrated Raman spectroscopy, fabricated using ALD assisted DUV lithography and therefore compatible with the back-end CMOS fabrication. We have reported both the experimental and numerical characterization of our sensor demonstrating capabilities that have so far been difficult to combine in a single sensor: (1) a large Raman conversion efficiency (≈1 × 10⁻⁹), thanks to simultaneous long interaction length, high confinement, and plasmonic enhancement (SMEF = 1.5 × 10⁷), (2) a low background due to the reduced overlap of the field with the core of the waveguide, (3) a good reproducibility of the Raman spectra (RSD < 5%) across different chips, thanks to the nanometer fabrication accuracy provided by atomic layer deposition, (4) a good tolerance to rather large input power due to a moderate local field enhancement, and (5) a broadband enhancement, thanks to the use of propagating plasmons rather than localized plasmon (nanoantennae). All these features make this kind of sensor ideal for the future integration with lasers, spectrometers, spectral filters, and detectors that will constitute a fully integrated Raman spectrometer chip. Given the compactness of our plasmonic sensor and the other elements, it is easy to foresee a chip capable of containing dozens or hundreds of sensors for performing multiplexed Raman sensing. Such an application is of great importance in biology,^51,52 especially the study of proteins accessible to small gaps.⁵³ Also, our choice of a monolayer as an analyte supports many bio-applications from clinical diagnostics, proteomics, genomics, and biomaterials to tissue engineering.^54–57 Moreover, our fabrication method may also open the way to other plasmonic-enabled applications that remained to date confined to the laboratory because of the need for e-beam lithography. This is the case for fast optical switches,^58,59 efficient frequency converters,⁶⁰ or optical nanoscopy.⁶¹ 

Si₃N₄ slot waveguides used for our experiments are fabricated on a 200 mm silicon wafer containing a stack of 2.3 ± 0.1 μm thick high-density plasma enhanced chemical vapor deposition (PECVD) silicon oxide SiO₂ and 220 nm thick PECVD Si₃N₄.⁶² The structures were patterned with 193 nm optical lithography and subsequently etched by fluorine based inductive coupled plasma-reactive ion-etch process to attain the final structure. The resulting average slot width is 150 nm. The waveguide topology is depicted in Fig. 6. Then, the structure is conformally coated with 58 nm of Al₂O₃ using atomic layer deposition (ALD). Due to a significant modal overlap with Al₂O₃, as shown in Fig. 6(b), the amount of Raman background from the Al₂O₃ layer is very low. In order to compare the Raman background generated by bulk Al₂O₃ and PECVD Si₃N₄, top-down Raman measurements are taken on two samples, i.e., 100 nm Al₂O₃/1 mm CaF₂ and 220 nm Si₃N₄/1 mm CaF₂. Both spectra are recorded using 10 mW pump power and 30 s integration time. The Raman spectra after thickness normalization are shown in Fig. 6(c). The result shows that Al₂O₃ exhibits minimal Raman background as compared to PECVD Si₃N₄. The ALD deposition is performed in a home built ALD setup that has a base pressure of 10⁻⁶ mbar. Deposition of the Al₂O₃ layer is performed using a thermal ALD process⁶³ at a substrate temperature of 120 °C using trimethylaluminum and water pulses of 5 s at 5 × 10⁻³ mbar. After the Al₂O₃ deposition, AZ 5214 resist is spin coated over the chip using 3500 RPM and 40 s spinning time. This is followed by an 11 s DUV exposure, 3 min post-bake at 120 °C, and a 50 s DUV flood exposure (for image reversal). After the DUV exposure, the development was performed in a solution containing AZ400:DI-water (3:1). A short O₂ plasma (PVA-TEPLA GIGAbatch 310 m, 6000 SCCM O2, 600 W, 750 mTorr) was performed to remove the residual resist. A thin 1-2 nm of Ti followed by a 13-15 nm of gold along the side wall is deposited using the Au-sputtering. Finally, a 100 μm gold waveguide is defined using the metal lift. In order to characterize the performance of our device, the sensor is functionalized with a monolayer of 4-nitrophenol that selectively binds to the gold. It is worth mentioning that in the future, the ALD coating step may become obsolete because of the use of extreme-UV sources for integrated photonics, as is currently happening in the nanoelectronics industry.

We use a standard binding recipe, e.g., chemisorption of thiols from an ethanolic solution,⁶⁴ to get a self-assembled monolayer of NTP. The procedure is as follows: the chips are cleaned thoroughly by first rinsing with acetone and isopropyl alcohol in an ultrasonic bath, followed by a 5 min exposure to an O₂ plasma. The chips are then immersed overnight into a beaker containing 1 mM NTP solution in ethanol. This is followed by a deep cleaning using pure ethanol and N₂ dry clean. As a result, a uniform monolayer of NTP is formed on top of the gold through gold-sulfur bond. Note that an overnight soaking to form an NTP monolayer is an overkill and this procedure can be accelerated (<1 h) by optimizing the thiol-ethanolic solution.^41,65 Similarly, the accessibility of an analyte to a narrow slot can also be enhanced using wetting techniques.⁶⁶ 

A commercial confocal Raman microscope (WITEC Alpha300R+) is used for coupling the light in and out of the chip, which is positioned vertically and end-fire coupled, as shown in Fig. 7. A 785 nm excitation diode laser (Toptica XTRA II) is used for excitation. The polarization of the excitation beam is set to the TE mode of the waveguide. A laser power of 1 mW measured before the entrance facet of a Zeiss 100×/0.9 EC Epiplan NEOFLUAR: ∞/0 objective is used. The scattered signal is collected in back reflection using the same objective and imaged on a 100 μm multimode fiber. This fiber functions as a confocal pinhole and entrance slit of the spectrometer. The spectrometer uses a 600 lines/mm grating to disperse the light onto a −73 °C cooled CCD camera (ANDOR iDus 401 BR-DD). In addition to that, an integrated sphere connected to a ThorLabs powermeter is used at the waveguide end facet to monitor the transmission. The objective and chip are aligned with 100 nm accuracy based on a maximum intensity of the waveguide Raman spectrum. Simultaneously, maximum light scattering along the waveguide is observed from a camera imaging the top surface of the chip.

Supplementary material consists of three different sections. The first section describes the derivation of η and ζ, presented in Eq. (1). The second section consists of plasmonic waveguide loss measurements. In the last section, the NTP Raman spectra measured from 24 different waveguides are reported.

The authors acknowledge Professor Pol Van Dorpe (imec) and Professor Nicolas Le Thomas (UGent) for useful discussion. We also acknowledge Liesbet Van Landschoot for taking the SEM images.

This research was funded by the ERC Grant InSpectra and FWO Belgium. Stéphane Clemmen thanks the F.R.S-FNRS for financial support. We also acknowledge the support of BOF (UGent, Belgium).